Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures

Area of Science:

• Solid-state physics
• Materials science
• Computational chemistry

Background:

• Optical properties are crucial for applications like solar panels and sensors.
• First-principles computation of these properties is computationally expensive and complex.
• Machine learning shows promise but lacks efficient methods for optical property prediction from crystal structures.

Purpose of the Study:

• Introduce GNNOpt, an equivariant graph neural network for predicting optical spectra.
• Enable high-quality optical property predictions with a limited dataset.
• Facilitate the screening of photovoltaic and quantum materials.

Main Methods:

• Developed GNNOpt, an equivariant graph neural network with universal embedding.
• Utilized Kramers-Krönig relations for predicting various optical properties.
• Trained the model on a dataset of 944 materials.
• Validated predictions with first-principles calculations.

Main Results:

• GNNOpt achieves high-quality predictions of optical properties, including absorption, dielectric function, refractive index, and reflectance.
• The model demonstrates excellent agreement with first-principles calculations for unseen materials.
• Successfully screened photovoltaic materials and identified potential quantum materials, like SiOs.

Conclusions:

• GNNOpt offers an efficient and accurate approach for predicting optical properties from crystal structures.
• The model accelerates the discovery of novel materials for diverse technological applications.
• Highlights the potential of graph neural networks in materials science and condensed matter physics.